]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * linux/kernel/softirq.c | |
3 | * | |
4 | * Copyright (C) 1992 Linus Torvalds | |
5 | * | |
6 | * Distribute under GPLv2. | |
7 | * | |
8 | * Rewritten. Old one was good in 2.2, but in 2.3 it was immoral. --ANK (990903) | |
9 | */ | |
10 | ||
11 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
12 | ||
13 | #include <linux/export.h> | |
14 | #include <linux/kernel_stat.h> | |
15 | #include <linux/interrupt.h> | |
16 | #include <linux/init.h> | |
17 | #include <linux/mm.h> | |
18 | #include <linux/notifier.h> | |
19 | #include <linux/percpu.h> | |
20 | #include <linux/cpu.h> | |
21 | #include <linux/freezer.h> | |
22 | #include <linux/kthread.h> | |
23 | #include <linux/rcupdate.h> | |
24 | #include <linux/ftrace.h> | |
25 | #include <linux/smp.h> | |
26 | #include <linux/smpboot.h> | |
27 | #include <linux/tick.h> | |
28 | #include <linux/irq.h> | |
29 | ||
30 | #define CREATE_TRACE_POINTS | |
31 | #include <trace/events/irq.h> | |
32 | ||
33 | /* | |
34 | - No shared variables, all the data are CPU local. | |
35 | - If a softirq needs serialization, let it serialize itself | |
36 | by its own spinlocks. | |
37 | - Even if softirq is serialized, only local cpu is marked for | |
38 | execution. Hence, we get something sort of weak cpu binding. | |
39 | Though it is still not clear, will it result in better locality | |
40 | or will not. | |
41 | ||
42 | Examples: | |
43 | - NET RX softirq. It is multithreaded and does not require | |
44 | any global serialization. | |
45 | - NET TX softirq. It kicks software netdevice queues, hence | |
46 | it is logically serialized per device, but this serialization | |
47 | is invisible to common code. | |
48 | - Tasklets: serialized wrt itself. | |
49 | */ | |
50 | ||
51 | #ifndef __ARCH_IRQ_STAT | |
52 | irq_cpustat_t irq_stat[NR_CPUS] ____cacheline_aligned; | |
53 | EXPORT_SYMBOL(irq_stat); | |
54 | #endif | |
55 | ||
56 | static struct softirq_action softirq_vec[NR_SOFTIRQS] __cacheline_aligned_in_smp; | |
57 | ||
58 | DEFINE_PER_CPU(struct task_struct *, ksoftirqd); | |
59 | ||
60 | const char * const softirq_to_name[NR_SOFTIRQS] = { | |
61 | "HI", "TIMER", "NET_TX", "NET_RX", "BLOCK", "IRQ_POLL", | |
62 | "TASKLET", "SCHED", "HRTIMER", "RCU" | |
63 | }; | |
64 | ||
65 | /* | |
66 | * we cannot loop indefinitely here to avoid userspace starvation, | |
67 | * but we also don't want to introduce a worst case 1/HZ latency | |
68 | * to the pending events, so lets the scheduler to balance | |
69 | * the softirq load for us. | |
70 | */ | |
71 | static void wakeup_softirqd(void) | |
72 | { | |
73 | /* Interrupts are disabled: no need to stop preemption */ | |
74 | struct task_struct *tsk = __this_cpu_read(ksoftirqd); | |
75 | ||
76 | if (tsk && tsk->state != TASK_RUNNING) | |
77 | wake_up_process(tsk); | |
78 | } | |
79 | ||
80 | /* | |
81 | * If ksoftirqd is scheduled, we do not want to process pending softirqs | |
82 | * right now. Let ksoftirqd handle this at its own rate, to get fairness, | |
83 | * unless we're doing some of the synchronous softirqs. | |
84 | */ | |
85 | #define SOFTIRQ_NOW_MASK ((1 << HI_SOFTIRQ) | (1 << TASKLET_SOFTIRQ)) | |
86 | static bool ksoftirqd_running(unsigned long pending) | |
87 | { | |
88 | struct task_struct *tsk = __this_cpu_read(ksoftirqd); | |
89 | ||
90 | if (pending & SOFTIRQ_NOW_MASK) | |
91 | return false; | |
92 | return tsk && (tsk->state == TASK_RUNNING); | |
93 | } | |
94 | ||
95 | /* | |
96 | * preempt_count and SOFTIRQ_OFFSET usage: | |
97 | * - preempt_count is changed by SOFTIRQ_OFFSET on entering or leaving | |
98 | * softirq processing. | |
99 | * - preempt_count is changed by SOFTIRQ_DISABLE_OFFSET (= 2 * SOFTIRQ_OFFSET) | |
100 | * on local_bh_disable or local_bh_enable. | |
101 | * This lets us distinguish between whether we are currently processing | |
102 | * softirq and whether we just have bh disabled. | |
103 | */ | |
104 | ||
105 | /* | |
106 | * This one is for softirq.c-internal use, | |
107 | * where hardirqs are disabled legitimately: | |
108 | */ | |
109 | #ifdef CONFIG_TRACE_IRQFLAGS | |
110 | void __local_bh_disable_ip(unsigned long ip, unsigned int cnt) | |
111 | { | |
112 | unsigned long flags; | |
113 | ||
114 | WARN_ON_ONCE(in_irq()); | |
115 | ||
116 | raw_local_irq_save(flags); | |
117 | /* | |
118 | * The preempt tracer hooks into preempt_count_add and will break | |
119 | * lockdep because it calls back into lockdep after SOFTIRQ_OFFSET | |
120 | * is set and before current->softirq_enabled is cleared. | |
121 | * We must manually increment preempt_count here and manually | |
122 | * call the trace_preempt_off later. | |
123 | */ | |
124 | __preempt_count_add(cnt); | |
125 | /* | |
126 | * Were softirqs turned off above: | |
127 | */ | |
128 | if (softirq_count() == (cnt & SOFTIRQ_MASK)) | |
129 | trace_softirqs_off(ip); | |
130 | raw_local_irq_restore(flags); | |
131 | ||
132 | if (preempt_count() == cnt) { | |
133 | #ifdef CONFIG_DEBUG_PREEMPT | |
134 | current->preempt_disable_ip = get_lock_parent_ip(); | |
135 | #endif | |
136 | trace_preempt_off(CALLER_ADDR0, get_lock_parent_ip()); | |
137 | } | |
138 | } | |
139 | EXPORT_SYMBOL(__local_bh_disable_ip); | |
140 | #endif /* CONFIG_TRACE_IRQFLAGS */ | |
141 | ||
142 | static void __local_bh_enable(unsigned int cnt) | |
143 | { | |
144 | lockdep_assert_irqs_disabled(); | |
145 | ||
146 | if (preempt_count() == cnt) | |
147 | trace_preempt_on(CALLER_ADDR0, get_lock_parent_ip()); | |
148 | ||
149 | if (softirq_count() == (cnt & SOFTIRQ_MASK)) | |
150 | trace_softirqs_on(_RET_IP_); | |
151 | ||
152 | __preempt_count_sub(cnt); | |
153 | } | |
154 | ||
155 | /* | |
156 | * Special-case - softirqs can safely be enabled in | |
157 | * cond_resched_softirq(), or by __do_softirq(), | |
158 | * without processing still-pending softirqs: | |
159 | */ | |
160 | void _local_bh_enable(void) | |
161 | { | |
162 | WARN_ON_ONCE(in_irq()); | |
163 | __local_bh_enable(SOFTIRQ_DISABLE_OFFSET); | |
164 | } | |
165 | EXPORT_SYMBOL(_local_bh_enable); | |
166 | ||
167 | void __local_bh_enable_ip(unsigned long ip, unsigned int cnt) | |
168 | { | |
169 | WARN_ON_ONCE(in_irq()); | |
170 | lockdep_assert_irqs_enabled(); | |
171 | #ifdef CONFIG_TRACE_IRQFLAGS | |
172 | local_irq_disable(); | |
173 | #endif | |
174 | /* | |
175 | * Are softirqs going to be turned on now: | |
176 | */ | |
177 | if (softirq_count() == SOFTIRQ_DISABLE_OFFSET) | |
178 | trace_softirqs_on(ip); | |
179 | /* | |
180 | * Keep preemption disabled until we are done with | |
181 | * softirq processing: | |
182 | */ | |
183 | preempt_count_sub(cnt - 1); | |
184 | ||
185 | if (unlikely(!in_interrupt() && local_softirq_pending())) { | |
186 | /* | |
187 | * Run softirq if any pending. And do it in its own stack | |
188 | * as we may be calling this deep in a task call stack already. | |
189 | */ | |
190 | do_softirq(); | |
191 | } | |
192 | ||
193 | preempt_count_dec(); | |
194 | #ifdef CONFIG_TRACE_IRQFLAGS | |
195 | local_irq_enable(); | |
196 | #endif | |
197 | preempt_check_resched(); | |
198 | } | |
199 | EXPORT_SYMBOL(__local_bh_enable_ip); | |
200 | ||
201 | /* | |
202 | * We restart softirq processing for at most MAX_SOFTIRQ_RESTART times, | |
203 | * but break the loop if need_resched() is set or after 2 ms. | |
204 | * The MAX_SOFTIRQ_TIME provides a nice upper bound in most cases, but in | |
205 | * certain cases, such as stop_machine(), jiffies may cease to | |
206 | * increment and so we need the MAX_SOFTIRQ_RESTART limit as | |
207 | * well to make sure we eventually return from this method. | |
208 | * | |
209 | * These limits have been established via experimentation. | |
210 | * The two things to balance is latency against fairness - | |
211 | * we want to handle softirqs as soon as possible, but they | |
212 | * should not be able to lock up the box. | |
213 | */ | |
214 | #define MAX_SOFTIRQ_TIME msecs_to_jiffies(2) | |
215 | #define MAX_SOFTIRQ_RESTART 10 | |
216 | ||
217 | #ifdef CONFIG_TRACE_IRQFLAGS | |
218 | /* | |
219 | * When we run softirqs from irq_exit() and thus on the hardirq stack we need | |
220 | * to keep the lockdep irq context tracking as tight as possible in order to | |
221 | * not miss-qualify lock contexts and miss possible deadlocks. | |
222 | */ | |
223 | ||
224 | static inline bool lockdep_softirq_start(void) | |
225 | { | |
226 | bool in_hardirq = false; | |
227 | ||
228 | if (trace_hardirq_context(current)) { | |
229 | in_hardirq = true; | |
230 | trace_hardirq_exit(); | |
231 | } | |
232 | ||
233 | lockdep_softirq_enter(); | |
234 | ||
235 | return in_hardirq; | |
236 | } | |
237 | ||
238 | static inline void lockdep_softirq_end(bool in_hardirq) | |
239 | { | |
240 | lockdep_softirq_exit(); | |
241 | ||
242 | if (in_hardirq) | |
243 | trace_hardirq_enter(); | |
244 | } | |
245 | #else | |
246 | static inline bool lockdep_softirq_start(void) { return false; } | |
247 | static inline void lockdep_softirq_end(bool in_hardirq) { } | |
248 | #endif | |
249 | ||
250 | asmlinkage __visible void __softirq_entry __do_softirq(void) | |
251 | { | |
252 | unsigned long end = jiffies + MAX_SOFTIRQ_TIME; | |
253 | unsigned long old_flags = current->flags; | |
254 | int max_restart = MAX_SOFTIRQ_RESTART; | |
255 | struct softirq_action *h; | |
256 | bool in_hardirq; | |
257 | __u32 pending; | |
258 | int softirq_bit; | |
259 | ||
260 | /* | |
261 | * Mask out PF_MEMALLOC s current task context is borrowed for the | |
262 | * softirq. A softirq handled such as network RX might set PF_MEMALLOC | |
263 | * again if the socket is related to swap | |
264 | */ | |
265 | current->flags &= ~PF_MEMALLOC; | |
266 | ||
267 | pending = local_softirq_pending(); | |
268 | account_irq_enter_time(current); | |
269 | ||
270 | __local_bh_disable_ip(_RET_IP_, SOFTIRQ_OFFSET); | |
271 | in_hardirq = lockdep_softirq_start(); | |
272 | ||
273 | restart: | |
274 | /* Reset the pending bitmask before enabling irqs */ | |
275 | set_softirq_pending(0); | |
276 | ||
277 | local_irq_enable(); | |
278 | ||
279 | h = softirq_vec; | |
280 | ||
281 | while ((softirq_bit = ffs(pending))) { | |
282 | unsigned int vec_nr; | |
283 | int prev_count; | |
284 | ||
285 | h += softirq_bit - 1; | |
286 | ||
287 | vec_nr = h - softirq_vec; | |
288 | prev_count = preempt_count(); | |
289 | ||
290 | kstat_incr_softirqs_this_cpu(vec_nr); | |
291 | ||
292 | trace_softirq_entry(vec_nr); | |
293 | h->action(h); | |
294 | trace_softirq_exit(vec_nr); | |
295 | if (unlikely(prev_count != preempt_count())) { | |
296 | pr_err("huh, entered softirq %u %s %p with preempt_count %08x, exited with %08x?\n", | |
297 | vec_nr, softirq_to_name[vec_nr], h->action, | |
298 | prev_count, preempt_count()); | |
299 | preempt_count_set(prev_count); | |
300 | } | |
301 | h++; | |
302 | pending >>= softirq_bit; | |
303 | } | |
304 | ||
305 | rcu_bh_qs(); | |
306 | local_irq_disable(); | |
307 | ||
308 | pending = local_softirq_pending(); | |
309 | if (pending) { | |
310 | if (time_before(jiffies, end) && !need_resched() && | |
311 | --max_restart) | |
312 | goto restart; | |
313 | ||
314 | wakeup_softirqd(); | |
315 | } | |
316 | ||
317 | lockdep_softirq_end(in_hardirq); | |
318 | account_irq_exit_time(current); | |
319 | __local_bh_enable(SOFTIRQ_OFFSET); | |
320 | WARN_ON_ONCE(in_interrupt()); | |
321 | current_restore_flags(old_flags, PF_MEMALLOC); | |
322 | } | |
323 | ||
324 | asmlinkage __visible void do_softirq(void) | |
325 | { | |
326 | __u32 pending; | |
327 | unsigned long flags; | |
328 | ||
329 | if (in_interrupt()) | |
330 | return; | |
331 | ||
332 | local_irq_save(flags); | |
333 | ||
334 | pending = local_softirq_pending(); | |
335 | ||
336 | if (pending && !ksoftirqd_running(pending)) | |
337 | do_softirq_own_stack(); | |
338 | ||
339 | local_irq_restore(flags); | |
340 | } | |
341 | ||
342 | /* | |
343 | * Enter an interrupt context. | |
344 | */ | |
345 | void irq_enter(void) | |
346 | { | |
347 | rcu_irq_enter(); | |
348 | if (is_idle_task(current) && !in_interrupt()) { | |
349 | /* | |
350 | * Prevent raise_softirq from needlessly waking up ksoftirqd | |
351 | * here, as softirq will be serviced on return from interrupt. | |
352 | */ | |
353 | local_bh_disable(); | |
354 | tick_irq_enter(); | |
355 | _local_bh_enable(); | |
356 | } | |
357 | ||
358 | __irq_enter(); | |
359 | } | |
360 | ||
361 | static inline void invoke_softirq(void) | |
362 | { | |
363 | if (ksoftirqd_running(local_softirq_pending())) | |
364 | return; | |
365 | ||
366 | if (!force_irqthreads) { | |
367 | #ifdef CONFIG_HAVE_IRQ_EXIT_ON_IRQ_STACK | |
368 | /* | |
369 | * We can safely execute softirq on the current stack if | |
370 | * it is the irq stack, because it should be near empty | |
371 | * at this stage. | |
372 | */ | |
373 | __do_softirq(); | |
374 | #else | |
375 | /* | |
376 | * Otherwise, irq_exit() is called on the task stack that can | |
377 | * be potentially deep already. So call softirq in its own stack | |
378 | * to prevent from any overrun. | |
379 | */ | |
380 | do_softirq_own_stack(); | |
381 | #endif | |
382 | } else { | |
383 | wakeup_softirqd(); | |
384 | } | |
385 | } | |
386 | ||
387 | static inline void tick_irq_exit(void) | |
388 | { | |
389 | #ifdef CONFIG_NO_HZ_COMMON | |
390 | int cpu = smp_processor_id(); | |
391 | ||
392 | /* Make sure that timer wheel updates are propagated */ | |
393 | if ((idle_cpu(cpu) && !need_resched()) || tick_nohz_full_cpu(cpu)) { | |
394 | if (!in_irq()) | |
395 | tick_nohz_irq_exit(); | |
396 | } | |
397 | #endif | |
398 | } | |
399 | ||
400 | /* | |
401 | * Exit an interrupt context. Process softirqs if needed and possible: | |
402 | */ | |
403 | void irq_exit(void) | |
404 | { | |
405 | #ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED | |
406 | local_irq_disable(); | |
407 | #else | |
408 | lockdep_assert_irqs_disabled(); | |
409 | #endif | |
410 | account_irq_exit_time(current); | |
411 | preempt_count_sub(HARDIRQ_OFFSET); | |
412 | if (!in_interrupt() && local_softirq_pending()) | |
413 | invoke_softirq(); | |
414 | ||
415 | tick_irq_exit(); | |
416 | rcu_irq_exit(); | |
417 | trace_hardirq_exit(); /* must be last! */ | |
418 | } | |
419 | ||
420 | /* | |
421 | * This function must run with irqs disabled! | |
422 | */ | |
423 | inline void raise_softirq_irqoff(unsigned int nr) | |
424 | { | |
425 | __raise_softirq_irqoff(nr); | |
426 | ||
427 | /* | |
428 | * If we're in an interrupt or softirq, we're done | |
429 | * (this also catches softirq-disabled code). We will | |
430 | * actually run the softirq once we return from | |
431 | * the irq or softirq. | |
432 | * | |
433 | * Otherwise we wake up ksoftirqd to make sure we | |
434 | * schedule the softirq soon. | |
435 | */ | |
436 | if (!in_interrupt()) | |
437 | wakeup_softirqd(); | |
438 | } | |
439 | ||
440 | void raise_softirq(unsigned int nr) | |
441 | { | |
442 | unsigned long flags; | |
443 | ||
444 | local_irq_save(flags); | |
445 | raise_softirq_irqoff(nr); | |
446 | local_irq_restore(flags); | |
447 | } | |
448 | ||
449 | void __raise_softirq_irqoff(unsigned int nr) | |
450 | { | |
451 | trace_softirq_raise(nr); | |
452 | or_softirq_pending(1UL << nr); | |
453 | } | |
454 | ||
455 | void open_softirq(int nr, void (*action)(struct softirq_action *)) | |
456 | { | |
457 | softirq_vec[nr].action = action; | |
458 | } | |
459 | ||
460 | /* | |
461 | * Tasklets | |
462 | */ | |
463 | struct tasklet_head { | |
464 | struct tasklet_struct *head; | |
465 | struct tasklet_struct **tail; | |
466 | }; | |
467 | ||
468 | static DEFINE_PER_CPU(struct tasklet_head, tasklet_vec); | |
469 | static DEFINE_PER_CPU(struct tasklet_head, tasklet_hi_vec); | |
470 | ||
471 | void __tasklet_schedule(struct tasklet_struct *t) | |
472 | { | |
473 | unsigned long flags; | |
474 | ||
475 | local_irq_save(flags); | |
476 | t->next = NULL; | |
477 | *__this_cpu_read(tasklet_vec.tail) = t; | |
478 | __this_cpu_write(tasklet_vec.tail, &(t->next)); | |
479 | raise_softirq_irqoff(TASKLET_SOFTIRQ); | |
480 | local_irq_restore(flags); | |
481 | } | |
482 | EXPORT_SYMBOL(__tasklet_schedule); | |
483 | ||
484 | void __tasklet_hi_schedule(struct tasklet_struct *t) | |
485 | { | |
486 | unsigned long flags; | |
487 | ||
488 | local_irq_save(flags); | |
489 | t->next = NULL; | |
490 | *__this_cpu_read(tasklet_hi_vec.tail) = t; | |
491 | __this_cpu_write(tasklet_hi_vec.tail, &(t->next)); | |
492 | raise_softirq_irqoff(HI_SOFTIRQ); | |
493 | local_irq_restore(flags); | |
494 | } | |
495 | EXPORT_SYMBOL(__tasklet_hi_schedule); | |
496 | ||
497 | static __latent_entropy void tasklet_action(struct softirq_action *a) | |
498 | { | |
499 | struct tasklet_struct *list; | |
500 | ||
501 | local_irq_disable(); | |
502 | list = __this_cpu_read(tasklet_vec.head); | |
503 | __this_cpu_write(tasklet_vec.head, NULL); | |
504 | __this_cpu_write(tasklet_vec.tail, this_cpu_ptr(&tasklet_vec.head)); | |
505 | local_irq_enable(); | |
506 | ||
507 | while (list) { | |
508 | struct tasklet_struct *t = list; | |
509 | ||
510 | list = list->next; | |
511 | ||
512 | if (tasklet_trylock(t)) { | |
513 | if (!atomic_read(&t->count)) { | |
514 | if (!test_and_clear_bit(TASKLET_STATE_SCHED, | |
515 | &t->state)) | |
516 | BUG(); | |
517 | t->func(t->data); | |
518 | tasklet_unlock(t); | |
519 | continue; | |
520 | } | |
521 | tasklet_unlock(t); | |
522 | } | |
523 | ||
524 | local_irq_disable(); | |
525 | t->next = NULL; | |
526 | *__this_cpu_read(tasklet_vec.tail) = t; | |
527 | __this_cpu_write(tasklet_vec.tail, &(t->next)); | |
528 | __raise_softirq_irqoff(TASKLET_SOFTIRQ); | |
529 | local_irq_enable(); | |
530 | } | |
531 | } | |
532 | ||
533 | static __latent_entropy void tasklet_hi_action(struct softirq_action *a) | |
534 | { | |
535 | struct tasklet_struct *list; | |
536 | ||
537 | local_irq_disable(); | |
538 | list = __this_cpu_read(tasklet_hi_vec.head); | |
539 | __this_cpu_write(tasklet_hi_vec.head, NULL); | |
540 | __this_cpu_write(tasklet_hi_vec.tail, this_cpu_ptr(&tasklet_hi_vec.head)); | |
541 | local_irq_enable(); | |
542 | ||
543 | while (list) { | |
544 | struct tasklet_struct *t = list; | |
545 | ||
546 | list = list->next; | |
547 | ||
548 | if (tasklet_trylock(t)) { | |
549 | if (!atomic_read(&t->count)) { | |
550 | if (!test_and_clear_bit(TASKLET_STATE_SCHED, | |
551 | &t->state)) | |
552 | BUG(); | |
553 | t->func(t->data); | |
554 | tasklet_unlock(t); | |
555 | continue; | |
556 | } | |
557 | tasklet_unlock(t); | |
558 | } | |
559 | ||
560 | local_irq_disable(); | |
561 | t->next = NULL; | |
562 | *__this_cpu_read(tasklet_hi_vec.tail) = t; | |
563 | __this_cpu_write(tasklet_hi_vec.tail, &(t->next)); | |
564 | __raise_softirq_irqoff(HI_SOFTIRQ); | |
565 | local_irq_enable(); | |
566 | } | |
567 | } | |
568 | ||
569 | void tasklet_init(struct tasklet_struct *t, | |
570 | void (*func)(unsigned long), unsigned long data) | |
571 | { | |
572 | t->next = NULL; | |
573 | t->state = 0; | |
574 | atomic_set(&t->count, 0); | |
575 | t->func = func; | |
576 | t->data = data; | |
577 | } | |
578 | EXPORT_SYMBOL(tasklet_init); | |
579 | ||
580 | void tasklet_kill(struct tasklet_struct *t) | |
581 | { | |
582 | if (in_interrupt()) | |
583 | pr_notice("Attempt to kill tasklet from interrupt\n"); | |
584 | ||
585 | while (test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) { | |
586 | do { | |
587 | yield(); | |
588 | } while (test_bit(TASKLET_STATE_SCHED, &t->state)); | |
589 | } | |
590 | tasklet_unlock_wait(t); | |
591 | clear_bit(TASKLET_STATE_SCHED, &t->state); | |
592 | } | |
593 | EXPORT_SYMBOL(tasklet_kill); | |
594 | ||
595 | /* | |
596 | * tasklet_hrtimer | |
597 | */ | |
598 | ||
599 | /* | |
600 | * The trampoline is called when the hrtimer expires. It schedules a tasklet | |
601 | * to run __tasklet_hrtimer_trampoline() which in turn will call the intended | |
602 | * hrtimer callback, but from softirq context. | |
603 | */ | |
604 | static enum hrtimer_restart __hrtimer_tasklet_trampoline(struct hrtimer *timer) | |
605 | { | |
606 | struct tasklet_hrtimer *ttimer = | |
607 | container_of(timer, struct tasklet_hrtimer, timer); | |
608 | ||
609 | tasklet_hi_schedule(&ttimer->tasklet); | |
610 | return HRTIMER_NORESTART; | |
611 | } | |
612 | ||
613 | /* | |
614 | * Helper function which calls the hrtimer callback from | |
615 | * tasklet/softirq context | |
616 | */ | |
617 | static void __tasklet_hrtimer_trampoline(unsigned long data) | |
618 | { | |
619 | struct tasklet_hrtimer *ttimer = (void *)data; | |
620 | enum hrtimer_restart restart; | |
621 | ||
622 | restart = ttimer->function(&ttimer->timer); | |
623 | if (restart != HRTIMER_NORESTART) | |
624 | hrtimer_restart(&ttimer->timer); | |
625 | } | |
626 | ||
627 | /** | |
628 | * tasklet_hrtimer_init - Init a tasklet/hrtimer combo for softirq callbacks | |
629 | * @ttimer: tasklet_hrtimer which is initialized | |
630 | * @function: hrtimer callback function which gets called from softirq context | |
631 | * @which_clock: clock id (CLOCK_MONOTONIC/CLOCK_REALTIME) | |
632 | * @mode: hrtimer mode (HRTIMER_MODE_ABS/HRTIMER_MODE_REL) | |
633 | */ | |
634 | void tasklet_hrtimer_init(struct tasklet_hrtimer *ttimer, | |
635 | enum hrtimer_restart (*function)(struct hrtimer *), | |
636 | clockid_t which_clock, enum hrtimer_mode mode) | |
637 | { | |
638 | hrtimer_init(&ttimer->timer, which_clock, mode); | |
639 | ttimer->timer.function = __hrtimer_tasklet_trampoline; | |
640 | tasklet_init(&ttimer->tasklet, __tasklet_hrtimer_trampoline, | |
641 | (unsigned long)ttimer); | |
642 | ttimer->function = function; | |
643 | } | |
644 | EXPORT_SYMBOL_GPL(tasklet_hrtimer_init); | |
645 | ||
646 | void __init softirq_init(void) | |
647 | { | |
648 | int cpu; | |
649 | ||
650 | for_each_possible_cpu(cpu) { | |
651 | per_cpu(tasklet_vec, cpu).tail = | |
652 | &per_cpu(tasklet_vec, cpu).head; | |
653 | per_cpu(tasklet_hi_vec, cpu).tail = | |
654 | &per_cpu(tasklet_hi_vec, cpu).head; | |
655 | } | |
656 | ||
657 | open_softirq(TASKLET_SOFTIRQ, tasklet_action); | |
658 | open_softirq(HI_SOFTIRQ, tasklet_hi_action); | |
659 | } | |
660 | ||
661 | static int ksoftirqd_should_run(unsigned int cpu) | |
662 | { | |
663 | return local_softirq_pending(); | |
664 | } | |
665 | ||
666 | static void run_ksoftirqd(unsigned int cpu) | |
667 | { | |
668 | local_irq_disable(); | |
669 | if (local_softirq_pending()) { | |
670 | /* | |
671 | * We can safely run softirq on inline stack, as we are not deep | |
672 | * in the task stack here. | |
673 | */ | |
674 | __do_softirq(); | |
675 | local_irq_enable(); | |
676 | cond_resched_rcu_qs(); | |
677 | return; | |
678 | } | |
679 | local_irq_enable(); | |
680 | } | |
681 | ||
682 | #ifdef CONFIG_HOTPLUG_CPU | |
683 | /* | |
684 | * tasklet_kill_immediate is called to remove a tasklet which can already be | |
685 | * scheduled for execution on @cpu. | |
686 | * | |
687 | * Unlike tasklet_kill, this function removes the tasklet | |
688 | * _immediately_, even if the tasklet is in TASKLET_STATE_SCHED state. | |
689 | * | |
690 | * When this function is called, @cpu must be in the CPU_DEAD state. | |
691 | */ | |
692 | void tasklet_kill_immediate(struct tasklet_struct *t, unsigned int cpu) | |
693 | { | |
694 | struct tasklet_struct **i; | |
695 | ||
696 | BUG_ON(cpu_online(cpu)); | |
697 | BUG_ON(test_bit(TASKLET_STATE_RUN, &t->state)); | |
698 | ||
699 | if (!test_bit(TASKLET_STATE_SCHED, &t->state)) | |
700 | return; | |
701 | ||
702 | /* CPU is dead, so no lock needed. */ | |
703 | for (i = &per_cpu(tasklet_vec, cpu).head; *i; i = &(*i)->next) { | |
704 | if (*i == t) { | |
705 | *i = t->next; | |
706 | /* If this was the tail element, move the tail ptr */ | |
707 | if (*i == NULL) | |
708 | per_cpu(tasklet_vec, cpu).tail = i; | |
709 | return; | |
710 | } | |
711 | } | |
712 | BUG(); | |
713 | } | |
714 | ||
715 | static int takeover_tasklets(unsigned int cpu) | |
716 | { | |
717 | /* CPU is dead, so no lock needed. */ | |
718 | local_irq_disable(); | |
719 | ||
720 | /* Find end, append list for that CPU. */ | |
721 | if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) { | |
722 | *__this_cpu_read(tasklet_vec.tail) = per_cpu(tasklet_vec, cpu).head; | |
723 | this_cpu_write(tasklet_vec.tail, per_cpu(tasklet_vec, cpu).tail); | |
724 | per_cpu(tasklet_vec, cpu).head = NULL; | |
725 | per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head; | |
726 | } | |
727 | raise_softirq_irqoff(TASKLET_SOFTIRQ); | |
728 | ||
729 | if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) { | |
730 | *__this_cpu_read(tasklet_hi_vec.tail) = per_cpu(tasklet_hi_vec, cpu).head; | |
731 | __this_cpu_write(tasklet_hi_vec.tail, per_cpu(tasklet_hi_vec, cpu).tail); | |
732 | per_cpu(tasklet_hi_vec, cpu).head = NULL; | |
733 | per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head; | |
734 | } | |
735 | raise_softirq_irqoff(HI_SOFTIRQ); | |
736 | ||
737 | local_irq_enable(); | |
738 | return 0; | |
739 | } | |
740 | #else | |
741 | #define takeover_tasklets NULL | |
742 | #endif /* CONFIG_HOTPLUG_CPU */ | |
743 | ||
744 | static struct smp_hotplug_thread softirq_threads = { | |
745 | .store = &ksoftirqd, | |
746 | .thread_should_run = ksoftirqd_should_run, | |
747 | .thread_fn = run_ksoftirqd, | |
748 | .thread_comm = "ksoftirqd/%u", | |
749 | }; | |
750 | ||
751 | static __init int spawn_ksoftirqd(void) | |
752 | { | |
753 | cpuhp_setup_state_nocalls(CPUHP_SOFTIRQ_DEAD, "softirq:dead", NULL, | |
754 | takeover_tasklets); | |
755 | BUG_ON(smpboot_register_percpu_thread(&softirq_threads)); | |
756 | ||
757 | return 0; | |
758 | } | |
759 | early_initcall(spawn_ksoftirqd); | |
760 | ||
761 | /* | |
762 | * [ These __weak aliases are kept in a separate compilation unit, so that | |
763 | * GCC does not inline them incorrectly. ] | |
764 | */ | |
765 | ||
766 | int __init __weak early_irq_init(void) | |
767 | { | |
768 | return 0; | |
769 | } | |
770 | ||
771 | int __init __weak arch_probe_nr_irqs(void) | |
772 | { | |
773 | return NR_IRQS_LEGACY; | |
774 | } | |
775 | ||
776 | int __init __weak arch_early_irq_init(void) | |
777 | { | |
778 | return 0; | |
779 | } | |
780 | ||
781 | unsigned int __weak arch_dynirq_lower_bound(unsigned int from) | |
782 | { | |
783 | return from; | |
784 | } |